The short answer: To benchmark a financial model, score it against a fixed library of best-practice criteria, weight each criterion by how much damage a failure would cause, and turn every low score into a specific, located, costed fix. EXL Cloud's Skill 21 — the Model Best Practice Benchmark — does exactly that with Claude, producing a scored assessment, a findings log and a phased remediation plan inside the workbook.
Why "is this model any good?" is so hard to answer
Ask three experienced modellers to review the same workbook and you'll get three different answers. One focuses on formatting. One hunts for hardcodes. One rebuilds the cash flow to see if it ties. All useful, none comparable — and none of them tell the owner what really matters: how risky is this model, and what would it cost to fix?
That's the gap a benchmark fills. Instead of opinions, you get a consistent rubric, a score you can track over time, and a remediation plan you can quote against.
What the benchmark actually measures
Skill 21 scores a workbook against a library of 53 criteria in 11 categories, each with a stable ID so findings can be tracked from run to run:
| Category | Criteria | What it looks at |
|---|---|---|
| Governance & Documentation (GOV) | 5 | Instructions sheet, change log, version control and file naming, cell legend, data source register |
| Structure & Flow (STR) | 7 | Tab order and section numbering, inputs/workings/outputs separated, master control, grouping not hiding, no legacy sheets, protection, header block with error bar |
| Inputs (INP) | 5 | Colour-coded inputs, centralised assumptions, data validation, no business hardcodes in formulas, no typed values in calculation areas |
| Formula Integrity (FRM) | 9 | No #REF! or error values, no external links or circularity, formula consistency, auditable complexity, volatile functions, names and table references, defined-name hygiene |
| Consistency (CON) | 1 | Repeating period sheets share one structure |
| Checks (CHK) | 4 | Validation sheet with an error count, checks left switched on, live failures resolved, tolerances held as inputs so checks can genuinely fail |
| Data & Feeds (DAT) | 6 | Raw data in tables, last refresh stamped, data stored once, helper tables sized for performance, mapping complete, queries documented |
| Outputs & Dashboards (OUT) | 6 | Outputs linked to calculations, dashboards present, pivots reconciled, robust cross-sheet links, error handling that doesn't hide problems, commentary that ties to the numbers |
| Formatting (FMT) | 5 | Body font, gridlines and headings, frozen headers, EXL Cloud palette, table naming |
| AI Governance (AI) | 2 | AI-built changes logged, AI-drafted commentary reviewed before use |
| 3-Way Integrity (TWI) | 3 | Balance sheet balances every period, closing cash ties, NPAT rolls into retained earnings |
Criteria that don't apply are marked not applicable rather than penalised. A pure reporting pack, for example, has no balance sheet engine, so the 3-way integrity criteria drop out.
For the thinking behind these standards, see The Implications of Not Following a Modelling Standard and Formula Construction Best Practice on the Model Citizn Knowledge Hub.
How scoring works
Every applicable criterion is scored 0 to 4, based on evidence from the workbook itself — sheet names, cell references, counts and proportions. Each criterion also carries a severity:
- Critical (weight 4) — failure could make a decision wrong. Balance sheet doesn't balance, checks don't cover the forecast.
- High (weight 3) — failure materially increases error risk.
- Medium (weight 2) — failure slows review or maintenance.
- Low (weight 1) — polish.
The headline is a severity-weighted score, with a separate score for Critical criteria so a well-formatted model can't hide a broken engine behind good presentation. Any Critical or High criterion scoring 2 or less becomes a finding.
One rule matters more than any other: score content, never existence. "There is an Instructions sheet" is not evidence. "The Instructions sheet has a six-step update procedure with cell references and a 40-entry glossary" is. A sheet that exists but says nothing useful doesn't earn the points.
What you get at the end
The benchmark adds seven clearly labelled BM tabs to the end of the workbook, wired together with BM_ named ranges and checked by their own validation sheet:
- BM Control — headline score, Critical score, number of findings, estimated fix hours, and a score by category.
- BM Direct Evidence Map — what was scanned, and tie-outs proving the evidence is complete.
- BM Assessment — every criterion, its score, the evidence and the gap.
- BM Findings Log — each finding with its location (sheet and range) and a specific recommendation.
- BM Remediation Plan — findings sequenced into dependency-ordered phases with hours.
- BM Data Plumbing — every data feed into the model and how it's refreshed.
- BM Validation Checks — the benchmark checks itself.
Fix in the right order
The remediation plan is phased so you never polish something you're about to rebuild:
- Baseline & governance — owner, purpose, version log, AI disclosure.
- Stop the bleeding — balance sheet balances every period, no broken references.
- Make the checks bite — checks cover every row and every period.
- Strip business hardcodes — every assumption moves to an input.
- Separate inputs and restyle — labelled input blocks, one-direction flow.
- Formula consistency — row-consistent formulas, no external links.
- Scenarios and outputs — a scenario switch, outputs that tie to the engine.
Because every task carries hours, the plan doubles as a scoping document. If you're an adviser, that's the difference between "this model needs work" and a fixed-fee proposal the client can approve.
Lite run or full run?
A Lite run is a direct scan of the open workbook by Claude — fast, and ideal for a first look or an internal health check. A full run adds evidence from a dedicated formula analysis tool and a deeper review of documentation content, and is the one to use when the score will be shown to a client or a board. Every run gets a Run ID and is snapshotted, so you can show the score improving as the remediation plan is worked through.
When to run a benchmark
- When you inherit a model from a previous adviser or employee.
- Before a model is used for a capital raise, acquisition, lending decision or board approval.
- Quarterly on models that are updated every month-end, to catch drift.
- After a major rebuild, to prove the rebuild worked.
Frequently asked questions
Does the benchmark change my model?
No. The model under review is treated as read-only. The benchmark adds its own BM tabs and leaves your formulas alone; fixing findings is a separate, deliberate step.
Can I benchmark a model that wasn't built in EXL Cloud?
Yes. The criteria are drawn from recognised industry standards, not just EXL Cloud conventions. Non-EXL models typically score lower on the EXL-specific criteria, which is useful information in itself.
How is this different from Skill 20?
Skill 20 asks "do these numbers make sense?" Skill 21 asks "is this model built well enough to trust?" A model can pass one and fail the other. Use both. Read about Skill 20 →
Full details, including example prompts and everything it adds to your workbook, are on the Skill 21 page.
The Model Best Practice Benchmark is Skill 21 in the EXL Cloud AI Skills Library, included with the Advanced Content plan. See all 20+ skills →